The major thrust of research, however, both in Australia and internationally, is the further development of coupled ocean-atmosphere climate models. These coupled models will be much more detailed than the relatively simple models used routinely since the mid-1980s. They are also much more expensive to run. However, rather than simply forecasting El Niņo behaviour (e.g., sea-surface temperatures in the east equatorial Pacific), these models could, in theory, be used to forecast rainfall and temperature over land. The representation of the coupling between the ocean and the atmosphere in these models, however, is less than perfect. CSIRO, with BMRC, is developing an improved ocean model (Power et al. 1995). This model, the Australian Community Ocean Model (ACOM), will be used in coupled model development.

One problem is the difficulty the atmospheric part of the coupled models has in simulating rainfall on the spatial scales important for users, despite their general success in simulating atmospheric variability (Nicholls 1996). Model experiments with specified sea-surface temperature anomalies have a long history in the Australian region (Simpson and Downey 1975; Voice and Hunt 1984). We have run our climate model, forced by observed sea-surface temperatures from 1949 to 1991, and compared rainfall simulated by the model with the observed rainfall (Frederiksen et al. 1995). We ran the model five times, with the same sea-surface temperatures. The only difference between the model runs was that slightly different starting atmospheric conditions were used. The difference between the model runs illustrates the 'noise' in the model. In fact, to get much agreement with observed rainfall, we need to average all five runs. This is 'ensemble' averaging. Our results indicate that we would need to run our models five or more times, with predicted sea-surface temperatures, or in coupled mode, and do routine ensemble averaging to reduce the inherent noise in the simulations. The ensemble averages of precipitation show some skill in simulating Australian precipitation, at least over northern Australia. In higher latitudes the models are less successful.

In these atmospheric model experiments, the model does very well in simulating the SOI. The SOI can, then, probably be predicted without the need to run expensive ensembles. If so, we could use the coupled ocean-atmosphere models to predict large-scale indices such as the SOI, then use statistical relationships between these forecast indices and the variables we really want to predict (e.g., rainfall at Emerald). Alternatively, we could use the coupled models to predict sea-surface temperatures, then use ensemble runs of an atmospheric model forced with predicted sea-surface temperatures to prepare predictions of rainfall and other variables of interest. Considerable thought and testing is needed to select the best strategy for using models in seasonal climate prediction.

Bureau of Meteorology